Lr_find() not plotting the same as in the Rossmann notebook

(Jacques Thibodeau) #1

(Posted this in Part 1, but this category may be better)

I’m currently going through the Rossmann (lesson 6) notebook and I’m running into some issues with lr_find().

Here’s what I’m supposed to get:

Here’s what I’m getting:

Notice I’ve even extended end_lr much farther. It cuts off at 1e+03 even though I used 1e+06.

I am using fastai version 1.0.58.

What could be causing this difference?

Here’s a GIST of my colab notebook: https://colab.research.google.com/gist/JayThibs/4ae331d045c5f6eeec49d09baa69e092/rossman-store-sales-tabular-data.ipynb

EDIT: I’m currently testing a few things. I used the same learning rate as Jeremy, and I’m getting a lower error (close to Jeremy’s but a little higher) than if I chose one based on the graph I received. Jeremy used a learning rate of 1e-03 and such a learning rate doesn’t make sense with the graph I received.

I re-ran the notebook and got a slightly different learning rate plot this time:

0 Likes